Object detection by using deep learning based approaches in remote sensing images

The issue of object detection in remote sensing images finds its place in many areas, especially in observations related to the earth. With the acquisition of richly detailed satellite and aerial images, it has become possible to detect and distinguish many man-made objects. However, variable object scales and appearances, as well as a number of display factors make this task challenging. On the other hand, advances in the field of deep learning and their adaptation to the field of remote sensing have provided significant improvements in the object detection task. In particular, region-based object detection tools using convolutional neural networks outperformed human performance in detection tasks in natural images. Mask R-CNN, which is a model in this structure, is one of the most up-to-date detection systems capable of extracting target objects in the given image with their masks. In this thesis, the use of the Mask R-CNN model for the detection of objects of this class in satellite images is proposed by considering the ship sample as a detection target. In order for the model to work at different scales with high accuracy, it is proposed to use the feature pyramid network, in which the upper layers of feature maps with high semantic values and the lower layers with high spatial resolution are fused. In addition, in order to reduce false alarms, training of the model with negative sampling or use of the focal loss function in the loss calculation is proposed. In order to perform a more quantitative error analysis, it is suggested that test images be classified according to their contents and the dimensions of the objects. Suggestions presented within the scope of this study were evaluated on the data sets created and showed the best performance in ship detection in terms of accuracy, precision and F1 scores.